SHRDLU for language understanding tries to find one good interpretation of an input. It doesn't maintain several possible interpretations of a sentence, it maintains one interpretation until it comes to an inconsistency and then backtracks.
This seems to work pretty well, and is at least a good heuristic approach, but I wonder if it gets in the way of language learning. In Cognitive Development, I talk about the necessity that each single input be used to train many parts of the brain. In the case of language processing, if we only allow one interpretation to be considered, we preclude correction across time. That is, while the single-interpretation approach might lead to best-possible performance within a single instant, in order to learn over time we need to have all possible interpretations trained by a given instant, so that over time we can learn to make better and better interpretations.